Professional Documents
Culture Documents
Keywords: continuous preference trend mining, product and design analytics, design for
multiple profit cycles, model tree
1 Introduction and Background Protection Agency (EPA) estimated that 2.37 106 short tons of
electronic products were ready for the end-of-life processes in
Product and design analytics relates and utilizes large-scale
2009 [6]. Compared to the amount in 1999, almost 50% have
data for design decisions throughout the life cycle of complex
increased and only 25% of them were gathered for recycling. The
products. The area of analytics is emerging as a promising area
second reason is the fact that electronic products contain toxic and
for the use of large-scale data generated by the users, original
hazardous materials [7]. Lead, mercury, nickel, and palladium are
equipment manufacturers (OEMs), markets, and the public, which
examples that present negative environmental and human health
are often freely available. Recent progress in this area has gar-
effects. Reckless landfills are not an optimal solution. A third fac-
nered notable achievements in predictive product design, for
tor of the problem of recovering these products is that electronic
example, merging predictive data mining and knowledge discov-
products also contain reusable and valuable resources, such as
ery methods with product portfolio design [1], predictive trend
gold, copper, tin, nickel, etc. [7]. Efficient and systematic methods
mining for product portfolio design [2], defining design analytics
to recover the reusable parts and resources are needed. Fourth,
[3], and linking on-line reviews with product attributes [4,5], to
more regulations and responsibilities are emerging. The countries
name a few. The area of product and design analytics could be
in the European Union have already begun adopting product take-
particularly beneficial for the area of product life cycle design and
back policy (Extended Producer Responsibility, EPR) since 1991
recovery, as product end-of-use state (e.g., life cycle length, prod-
[8]. The U.S. has also introduced more EPR laws recently com-
uct condition, product recovery decision making) is hard to pre-
pared to 2006 [9]. Fifth, “green consumers” [10] give more pres-
dict. The current paper shows that product and design analytics
sure to companies regarding their “green” image. Now, their
can serve as a foundation to link product prelife (design and man-
increased awareness of sustainability is a critical factor in deter-
ufacturing), usage, and end-of-life operations with the develop-
mining the demand of target products. Lastly, product recovery
ment of the trend mining algorithm.
and recycling are known to reduce the necessity for fuel consump-
tion and landfill space, and provide substantial benefits environ-
1.1 Recovery of End-Of-Life Electronics. Recovery of end- mentally [11].
of-life electronic products has become an urgent problem that Some OEMs such as Caterpillar, Xerox, and Sony have shown
requires design engineers’ attention due to multiple reasons. The first that a proper recovery system of their end-of-life products not
reason is the fast growing e-waste stream. The U.S. Environmental only extends their products’ lives and gives some environmental
benefits, but also allows for multiple profit cycles [12–14]. These
1
Corresponding author. OEMs consider the end-of-life stage as the “relife” stage
Contributed by the Design Automation Committee of ASME for publication in
the JOURNAL OF MECHANICAL DESIGN. Manuscript received May 23, 2013; final
and return take-back components to “same-as-new” condition
manuscript received February 14, 2014; published online April 11, 2014. Assoc. to customers. The relife processes or recovery options include
Editor: Bernard Yannou. reuse, repair, refurbishment, cannibalization, and recycling. By
Mean absolute error (MAE) 39.6 37.4 35.4 34.8 32.8 33.0 33.4 33.7 34.7 35.5 37.0
Total number of instances 9000 6125 4321 3199 2477 1998 1666 1427 1250 1111 1000
attributes and class variables. Then, a model tree can be built models classify class values without data so that no comparison
based on this data and tested with t ¼ 10 data. The best a is 0.4 can be made between a node and subtree for the pruning. In
based on the performance measure, mean absolute error. Sec. 3, pruning in the model tree algorithm will be introduced.
Based on the selected a, a number of required instances for
each data set is determined and sampled as in the example but
using t ¼ 1 to t ¼ 10 data sets this time. By applying predicted val- 2.1.3 Model Tree Induction. The third step is to build a model
ues of attributes and class variables at a target time to the sampled based on the newly generated data set from the second step. In
normalized data, trend embedded future data is finally generated this step, the knowledge and hidden patterns between the new val-
at the target time. Table 2 shows that the total number of samples ues of attributes and class variables are mined using a model tree.
can be varied depending on the selected a. Based on the smooth- The result of the model tree is a piecewise linear regression equa-
ing factor, only the latest data or all data can be used in the tion depending on a given data set, which can approximate nonlin-
extreme case. ear functions. Figure 5 shows an example of a model tree. The
A graphical example of the trend embedded future data genera- model tree gives three different linear models to express the nonli-
tion is depicted in Fig. 4. The values of original data sets are nor- nearity with two attributes: A and B. On the other hand, a decision
malized within a single time step and then sampled using the tree that was used in the other trend mining algorithms classifies
geometric sampling method with the selected a. In the example, discrete or categorical class variables.
suppose that the first and the last instances were sampled from the The M5 model tree was initially proposed by Quinlan [38].
t ¼ 10 data set. By applying predicted minimum and maximum After comprehensive descriptions of model tree induction includ-
values from the time series prediction technique, real values are ing a pseudocode by Wang and Witten [39], the model tree has
predicted at the target time. For example, the display size is get-
ting bigger, and the generated target data set reflects the trend
(e.g., refer to the small arrows).
By generating future data, two advantages can be achieved.
First, performance information of built models can be provided
similar to normal data mining processes. The predicted Gain Ratio
based models in Sec. 1.2 cannot give test data but the generated
data from the CPTM can work as test and validation data. The 10-
fold cross-validation technique [28] is a popular way to get the
performance (i.e., prediction of class variables) information when
a validation data is not available. In the 10-fold cross-validation,
the generated data are randomly partitioned into 10 subsamples
and validation processes are repeated 10 times. Each time a model
is built using 9 subsamples and validated with one remaining sub-
sample. Then, an average performance error can be estimated.
Second, pruning can be implemented based on the generated data
to reduce the risk of over-fitting. The predicted Gain Ratio based Fig. 5 Example of model tree
Fig. 7 Data from stationary linear mapping function and generated future data
Display Weight Hard CPU Graphics card Memory Battery Touchscreen Price
size (inch) (lbs) drive (GB) (technology) (technology) (GB) (hours) (technology) ($)
preprocessing the original data, monthly data sets of 155 cell with real t ¼ 12 data set in terms of the MAE. This procedure con-
phones from June 2009 to March 2010 were tested with 10 differ- tinued up to t ¼ 24 (total 13 times) for one time-ahead prediction
ent attributes: camera pixel, talk time, touch screen, weight, mem- and the results are shown in Fig. 9.
ory slot, wiFi, MP3, GPS, bluetooth, and 3G. In order to obtain a statistically valid conclusion between static
The goal is to construct a predictive model for the t ¼ 10 data models and the CPTM, both parametric (F and T-test) and non-
with t ¼ 1 to t ¼ 9 data sets. The smoothing factor, a ¼ 0, was parametric (Mann-Whitney) tests were employed. With a signifi-
selected using the built model from t ¼ 1 to t ¼ 8 data and tested cance level of alpha ¼ 0.05, the accuracy of the CPTM model was
with t ¼ 9 data. Also the automatic time series forecasting was significantly higher than that of static models with the generation
conducted in Table 4. Table 5 shows the result of performance of trend embedded data.
test with this real data.
3.5 Discussion. The CPTM algorithm showed good predic-
tive performances in comparison to the four well-known static
3.4 Test With Data Generated From Nonstationary Linear models in Table 5. From the cases of data sets generated from
Mapping Function. Twenty four data sets were generated ran- simple stationary linear and nonlinear mapping functions, it is rel-
domly with assumed ranges of attributes and some trends reflect- atively clear to look at the effect of the geometric sampling and
ing real-world tablet PC leasing markets. Each data set has 200 the time series prediction of attributes and class variables as
instances and 8 different attributes shown in Table 6. The first part shown in Figs. 7, 8, and Table 4. The geometric sampling helped
of Table 6 explains levels of each attribute which are decision var- to reflect the trend of relation between attributes and class varia-
iables explored in Sec. 2.2.2. The second part indicates an exam- bles over time. The automatic time series forecasting also gave
ple of the generated data. The data set shows transactional history good approximations of future attribute and class values. In both
and the class variable is the price that customers paid. Since this is cases, smoothing factors were close to zero, which makes sense in
data from nonstationary linear mapping functions, mapping func- that stationary mapping functions were applied over time.
tions with some randomness vary over time. The real data in Sec. 3.3 had an interesting data structure. The
The goal is to construct a series of predictive models for the values of attributes were fixed but the class variable continued to
t ¼ n þ 1 data using t ¼ 1 to t ¼ n data sets. n were increased by 1 change, which is why there are only predictions for class variables
from 11 to 23. For example, for an unseen t ¼ 12 data set, static in Table 4. It is important to realize that even though the forecast
models were constructed using the latest data set t ¼ 11 while the result was similar to the latest data, the geometric sampling
CPTM mined a trend from t ¼ 1 to t ¼ 11 data sets and constructed improved the predictive performance alone. Empirical tests showed
a tree model from a predicted data set at t0 ¼ 12 with the calcu- that without a precise prediction of attribute values, the CPTM
lated smoothing factor 0.5. Then, both models were evaluated algorithm worked well with the geometric sampling. Also, without
any knowledge of customer preferences and their trend over time in
the used product market, the selected smoothing factor can indicate
that underlying relations between attributes and class variables
were quite stationary in the interval of one month. This work with
real data has great potential and can provide some directions from
data collection to real application in different design domains.
The last case of data from nonstationary mapping functions rep-
resents a very complex data structure, and the question was
whether the CPTM worked well in this case. Due to the nonsta-
tionary nature of data, the prediction error of the CPTM was close
to other static models, e.g., at t ¼ 13 and t ¼ 14, etc., in Fig. 9.
However, statistical results showed that an overall performance of
the CPTM is better than other models with this data.
Among those four static models, the model tree was selected for
the purpose of direct comparison since the CPTM algorithm also
adopts the model tree for the prediction of class variables. From all
the tested data cases, the predictive performance of the CPTM out-
performed that of the model tree, and this indicates that the genera-
tion of a trend embedded data set improved the accuracy.
Fig. 9 Comparison of one time-ahead prediction accuracy While conducting these experiments, a total of five possible
between static and dynamic model sources of variation on the result were observed: smoothing
Objective function
1 1 1
Maximize f ¼ ðp12 c12 Þ þ ðp13 c13 Þ þ ðp14 c14 Þ þ ðp15 c15 Þ
ð1:03Þ0:5 ð1:03Þ ð1:03Þ1:5
p12 ¼ MT 12 ðYij ; Xij Þ, p13 ¼ MT 13 ðYij ; Xij Þ; p14 ¼ MT 14 ðYij ; Xij Þ; p15 ¼ MT 15 ðYij ; Xij Þ
X manufacturing
c12 ¼ cj Xij þ cforwardlogistics
j
c13 ¼ cinverselogistics þ cEOLð13Þ þ cforwardlogistics
c14 ¼ cinverselogistics þ cEOLð14Þ þ cforwardlogistics
c15 ¼ cinverselogistics þ cEOLð15Þ þ cforwardlogistics
Constraints
X
h1 : Yij ¼ 1
j
X
h2 : Xij ¼ 1
j
X manufacturing
h3 : cEOLðtÞ ¼ cj Xij ð1 bÞ þ creconditioning b
Q X j
h4 : b ¼ cj Xij
i j
h5 : Yij ; Xij 2 ð0; 1Þ
Display size Weight Hard drive CPU Graphics card Memory Battery Touch screen
(inch) (lbs) (GB) (technology) (technology) (GB) (hours) (technology)
C A midpoint stdev(T) stdev(T1) stdev(T2) SDR t ¼ 11 Hard drive CPU Graphics card Memory Battery Touch screen
t ¼ 15 Hard drive CPU Graphics card Memory Battery Touch screen Hard drive CPU Graphics card Memory Battery Touch screen
9 8 6 7 5 5 40 60 90 20 40 50
8 7 5 6 2 4 55 70 110 30 50 60
7 6 4 5 1 3 75 90 130 40 60 85
6 4 3 4 0 2 90 100 140 80 70 100
5 3 2 3 100 110 160 135
4 2 120 125